measuring coworker effects: evidence from an american ... · coworkers. productivity can be...

Measuring Coworker Effects:

Evidence from an American Manufacturing Firm

Jacob Adams

Advised by Takao Kato

Colgate University, 2018

This paper studies coworker effects and the dynamic of social identity, determined by

worker tenure, in a manufacturing setting. Using a unique data set to measure time invariant

productivity, or ability, of each worker, I find evidence of a free-rider effect stemming from an

increase in average ability of similarly tenured coworkers. The lack of performance-based pay

structures incentivizes workers to exert lower levels of effort when working with high ability

coworkers.

JEL Codes: D900, J000, L610

Keywords: Free-Riding, Peer Effects, Coworker Effects, Social Identity

2

1. Introduction

In many settings, a worker’s productivity is at the mercy of his or her coworkers because

workers are jointly responsible in the production process. For example, a single construction

worker can accomplish little individually, and the pace at which work is completed depends on

the productivity of the group. Consequently, measuring changes in individual coworker

productivity becomes very challenging. Mas and Moretti (2009) and Kato and Shu (2016) avoid

this problem by investigating empirical settings in which workers share a workspace, but the

production process depends on the solo work of many individuals rather than a single group.

Those papers study grocery store cashiers and Chinese textile workers, respectively. This paper

approaches the issue in a similar fashion by studying an American copper wire manufacturing

firm.

This paper finds its place in the literature by examining a manufacturing environment in

which there are no incentive pay structures and an unexplored social identifier. Additionally, I

investigate the significance of changes in team size on the production of the individual. Social

identity is not exogenously assigned, but rather, it depends on the distribution of worker tenure in

the firm, which creates two groups: temporary and committed workers. The workers all operate

machines in various stages of the production process of copper wire, but an individual’s

efficiency is independent of the efficiency of his or her peers. The firm does not use piece rate or

relative wage structures, but instead, a flat hourly wage independent of a worker’s efficiency or

relative efficiency.

By predicting the time invariant productivity (ability) for each worker in the firm, I

estimated the effects of changing the average worker’s ability in a work group. The results show

no evidence of an overall coworker effect – an increase in the average ability of all coworkers is

3

not associated with a change in a focal worker’s efficiency. However, there is a strong effect

stemming from in-group, similarly tenured, coworkers. An increase in the average ability of in-

group coworkers is associated with a decrease in a focal worker’s efficiency. This effect is very

consistent across the studied population; there is no evidence that this effect is any different for a

certain group of coworkers, or that the effect is non-linear. In addition, there is no evidence that

an overall increase in the number of workers in a work group increases a focal worker’s

efficiency, which is intriguing, as a higher number of coworkers could imply a higher level of

social pressure to perform well. In the absence of incentive pay structures, I attribute the negative

effects to an incentive to free ride off the work of more talented coworkers. The rationale for the

effect being exclusive to in-group coworkers lies in the empirical setting; the social dynamic

associated with tenure groups makes it easier for coworkers to build an understanding of the

innate ability of coworkers with whom they socialize.

2. Literature Review & Background:

Research on peer effects in the workplace has suggested that individuals adjust their level

of output and productivity according to the ability and/or various traits of their surrounding

coworkers. Productivity can be adjusted on account of a coworker’s differing social identity

(Kato and Shu 2016), whether a coworker is a friend (Bandiera, Barankay, and Rasul 2010), and

even whether or not an individual is in the sightline of a very productive worker (Mas and

Moretti 2009). Despite a high level of influence in very contextualized situations, the lack of

external validity remains a common thread in many empirical studies on peer effects in the

workplace.

Kato and Shu (2016) is perhaps the most relevant to this study because of the

manufacturing setting, but there are key differences. Kato and Shu attribute coworker effects to

4

an exogenously assigned variable: Hukou status. Hukou serves as a system of household

registration that provides those with “Urban Hukou Status” a higher level of economic

opportunity. As such, social identity theory plays a significant role in explaining why urban

Hukou holders increase their productivity significantly when the average ability of rural

coworkers increases: urban workers hold rural workers in contempt and perceive them as

inferior. Additionally, the Chinese textile firm in Kato and Shu utilizes a relative wage scale.

Weavers are pitted against one another, with the highest performers receiving the highest wages,

creating motivation to perform better than one’s peers. Similarly, the empirical setting of Mas

and Moretti (2009) allows for the study of peer effects as they pertain to the presence of high

ability workers causing a productivity boost in a work setting (in this case, a large chain grocery

store checkout). However, the effect is subject to a focal worker’s orientation to the highly

productive coworker in the work space. The grocery store cashiers are shown to be increasingly

sensitive to the presence of a highly achieving coworker the closer their checkout lines are to one

another. For this paper, the empirical setting is a bare copper wire manufacturing firm in Upstate

New York. Much like Kato and Shu (2016), the workers in this empirical setting are responsible

for operating machines, but unlike the Chinese textile firm, workers vary in the type of machine

they use and products they produce. Peer effects in settings like grocery bagging and

manufacturing are also shown to be stronger and easier to identify since the labor is relatively

low skilled - Cornelissen, Dustmann, and Schoenberg (2013) found evidence that as the skill

required for a job increases, the strength of peer effects decreases.

Comparing competition structures between peers, Mas and Moretti (2009) most closely

matches this paper; there is no incentive to work above what is expected. Wages in both settings

are not determined by production, performance, or experience. In Kato and Shu (2016) and

5

Bandiera, Barankay, and Rasul (2010), employees are working in systems that reward higher

productivity with higher pay. Chinese textile workers are paid based on their ability to produce a

lower level of defective output, while fruit pickers are paid a piece rate per a kilogram of fruit

picked. From this, it can be hypothesized that performing above expected efficiency can be

caused by either social pressure (as in Mas and Moretti 2009) or prosocial/altruistic behavior.

Nonetheless, highly productive fruit pickers in Bandiera, Barankay, and Rasul (2010) show a

willingness to forego income to work at a lower social norm. This is a surprising outcome, as

workers are shown to value their social standing more than potential economic gains.

Alternatively, it is possible the effort put forward is a result of a focal worker’s view on the

fairness of his/her wage; it has been suggested that productivity can be boosted by a wage

increase if the original wage was perceived to be unfairly low – Cohn, Fehr, and Goette (2014)

found evidence of this phenomenon by hiring workers for a one-time job in Switzerland,

studying their productivity, and surveying the workers on perceived fairness of pay.

Methodologically, Kato and Shu (2016) has strong influence on this paper. Using the

panel data, time-invariant productivity is estimated, and the lack of systematic department

assignment based on ability is confirmed. Time invariant productivity, regarded as ability, is

estimated by a productivity measure while controlling for the presence of other coworkers,

worker fixed effects, and time. Furthermore, the effects of productivity spillover based on spatial

orientation, as in Mas and Moretti (2009), is achieved by using data describing which workers

were present from a given department on a given shift. Finally, another potentially confounding

factor to consider is a superstar effect. It has been suggested through research of professional

golf tournaments that the presence of Tiger Woods, a highly skilled player, causes other players

to perform worse on average, compared to when the superstar is absent (Brown 2011). This

6

conflicts with the effect found in Mas and Moretti (2009), which is a positive superstar effect,

meaning that the work environment and different incentives have a strong level of influence on

the sign and magnitude of peer effects. However, it is important to note that the source of the

negative effects in golf tournaments is the lower motivation to perform because the odds of

winning the tournament are lowered by the presence of a superstar. Cashiers and machine

operators are in much different situations because they do not compete with their coworkers for a

payout.

Some studies associated with peer effects also take social identity into account. While the

empirical setting of this paper lacks an exogenously assigned social identity, the uneven

distribution in tenure among the workers opens the possibility of social group creation. The lack

of exogenously assigned social identity may reflect typical human behavior better than an

exogenously assigned social identifier – social groups are rarely determined by a single factor,

yet much of the social identity literature studies these uncommon situations. Moreover, it can

become very difficult to quantify social connections without a survey, but even then, there is

room for confounding factors - Cohn, Fehr, and Goette (2014) suggested that variation in the

time at which their survey was completed by subjects may have influenced their results.

While there is no potential to reward or punish coworkers for acting in line with or

against social norms in the copper wire firm, small, seemingly insignificant differences, like

tastes in art, can cause the formation of social alliances and influence behavior towards out of

group members. Chen and Li (2009) found that randomly assigned groups and groups

determined by art preference both exhibit behavior that differs significantly when interacting

with out-group players compared to in-group players. Specifically, they found matching with an

in-group member is associated with a 47% increase in charity concerns and a 93% decrease in

7

envy (Chen and Li 2009). Similarly, strategic behavior can be greatly influenced when group

membership is salient. Rigotti and Rustichini (2007) found that having more group members

present causes more coordinated group effort in search of a higher payoff in the “Battle of the

Sexes” game but causes a higher likelihood of defection in the Prisoner’s Dilemma. Bernhard,

Fehr, and Fischbacher (2006) performed a similar lab experiment by having two distinct tribes in

Papua New Guinea play the dictator game. They found a third party is more willing to punish

norm violators if the victim is in the same group and that compliance is stronger within groups

than between groups.

A secondary aim of this paper is to investigate the effect of a change in the size of a focal

worker’s team. The literature on peer effects stemming from this change is limited, but Haleblian

& Finkelstein (1993) offers some insight, as they show that top management team size has

positive effects on firm performance; firms with CEO dominated management were shown to

perform significantly worse than firms with large leadership teams. Given the manufacturing

setting of this paper and the fact workers are not managers, this paper adopts a unique

perspective. Workers in the manufacturing firm are assigned to their departments randomly,

meaning that they are put in positions that do not necessarily lend themselves to the workers’

strengths. This is much different from a large corporate leadership team that likely assigns

positions based on the skills of the individuals.

Finally, there are other issues concerning the study of peer effects in complicated

empirical situations that can possibly bias or interfere with the results. The potential for

promotion is oftentimes a difficult concept to quantify but is considerably important in the case

of a manufacturing plant; it has been suggested that helping efforts are reduced and individual

efforts are raised considerably when promotion incentives are strong (Drago & Garvey 1998).

8

The empirical situation of this paper does not contain high promotion incentives, so it is unlikely

that this style of incentive influences the behavior of workers in any significant manner.

This paper aims to expand the peer effects literature by offering evidence of coworker

effects through the interaction between worker efficiency and the average ability of coworkers

while accounting for social group membership, determined by tenure. In this case, tenure creates

a social identity that is not as strictly defined as hukou status or tribe membership (Bernhard,

Fehr, and Fischbacher 2006). Rather, it is more likely tenure variations create bonds between

workers much like the friendships shown in Bandiera, Barankay, and Rasul (2010). Considering

the uncommon pay structure of the manufacturing firm and the presence of a social identifier that

has never been explored, this paper investigates effects that demonstrate the variety of dynamics

that influence human interaction in an economic setting.

3. Empirical Setting:

The American manufacturing firm in this study is headquartered in Upstate New York,

with a total of 11 plants in various locations across the United States. The firm manufactures

many forms of bare copper wire from raw copper rods. The three plants examined in this paper

operate in Upstate New York. In collaboration with the firm, I obtained personnel and

productivity data that includes personal characteristics, daily performance measures, wages, and

production information. The observations take place over an eight-month span between January

and August 2017. The raw dataset contains a total of 131,977 observations spread across a total

of 223 workers and three plants. One observation comprises of data taken on one employee on a

single machine for one continuous period, typically the length of one shift.

9

3.1 Efficiency Metric

There are a variety of processes that go into production; workers use several machines

and produce a variety of products depending on which department they were assigned into when

hired. The machines are not automated, and output is heavily dependent on the skill and effort of

the worker. The typical role of a worker is to load his or her machine with the necessary inputs,

start the machine, and monitor the production process by dealing with potential breakdowns and

errors.

The performance metric in this study is efficiency. Efficiency is calculated by the firm

based on what the engineers believe to be the maximum safe output for a given machine over a

set period of time. Workers are expected to run their machines at 80 percent capacity for the

length of their shifts. So, if a machine can produce an output of X at 100 percent capacity for the

length of a shift, the standard that a worker is expected to produce is .80X. This level of output

would return a worker efficiency rating of 100 percent for that shift. Therefore, it is possible for

a skilled worker to receive an efficiency rating over 100 percent by operating his or her machine

at greater than 80 percent capacity. Workers typically run four or five separate machines each

shift, so the daily efficiency rating is calculated using the average. The time spent on each

machine depends on an individual worker’s ability to complete an order. Moving away from the

assigned work schedule should not be done without consulting a supervisor. Variation in the

products produced by workers is not a significant factor influencing efficiency, as most products

are made using the same set of machines. Thus, changes in the demand for certain products have

negligible effects on the daily work of an individual. Since efficiency is monitored by each

machine’s computer, workers are informed of their efficiency following every shift. Workers are

10

aware of the expected standard and know that consistent low efficiency can be considered

grounds for dismissal. The summary statistics for average efficiency are described in Table 7.

Furthermore, wage is based on what job a worker does on the factory floor, or more

specifically, what machines a worker runs. Wages range from $14.5 to $18 per hour. The wage

data summary statistics can be found in Table 7 along with summary statistics for employee age

and tenure. Because wages are fixed, there is no tangible incentive to work beyond the expected

80 percent production standard. There is also no incentive for promotion, as working in a

managerial role requires a level of knowledge and education that is not typically held by factory

floor workers. Moving to a higher paying job is not regarded as a promotion, but rather a lateral

move, because higher paying positions on the factory floor typically involve more physically

stressful work with no increase in responsibility. Workers may potentially move to a more

demanding job in search of a higher hourly wage, but this depends on the creation of an open

spot in a higher paying department. The only true motivational factor in the work setting is the

risk of being fired if the managers believe low efficiency is being caused by excessive laziness or

free-riding. Because of the lack of tangible incentives to work harder, workers may decide to

work harder based on either social pressure or prosocial behavior (Mas and Moretti 2009).

3.2 Tenure Differences

An important dynamic of the manufacturing firm in this study is the presence of a large

divide in worker tenure. As shown by Figure 1, there is a significant drop off in the number of

workers with over four years of tenure. Because of this, most new hires are not expected to be

retained for most of their working lives. This distribution creates a separation of “committed”

and “uncommitted” (called “temporary” in this study) workers. While there may be workers in

the “uncommitted/temporary” group that end up working for the firm for an extended period,

11

they have yet to show it. Furthermore, the distribution of age is shown in Figure 2. Age is much

more evenly distributed than tenure, suggesting that tenure is not strictly determined by age. That

being said, long tenured workers have to be of a certain age, and there exists generational

differences between workers in their 50’s and 60’s versus those in their 20’s and 30’s that likely

determine some social bonds in the workplace.

Long tenured, committed workers have reason to act differently around newer workers

because new hires must be trained and taught all the intricacies of production on the factory floor

while also learning to navigate the social landscape of the work place during breaks throughout

the shift. Workers tend to associate with coworkers they are more familiar with, and tenure often

determines with whom a worker is acquainted for a variety of reasons. The workers may have

been trained at the same time, have mutual friends in the workplace, or share common interests.

Longer tenured workers also do not expect that most of the training will be a long-term

investment since most workers end up leaving within four years. The effort and behavior of less

experienced workers also varies and may sway a longer tenured worker’s opinion. Moreover,

new hires are easy to spot in the workplace because they are required to wear a high visibility

vest for the first four weeks of work.

Due to the general social divide caused by large tenure differences among the firm’s

employee population, there is good reason to believe that long tenured workers behave

differently around coworkers with similar tenure compared to coworkers with vastly less

experience. For example, a long tenured work could potentially resent a new hire because even

though the tenured worker has a significantly higher level of skill development, he or she is still

paid the same wage as the new hire. Although, both temporary and uncommitted workers may

feel reason to behave differently around others in their same tenure group; the effect could

12

potentially vary in the directional effect on efficiency. For example, if two long tenured workers

are familiar with one another, they may revert to an efficiency norm (which may be higher or

lower that the focal worker’s time invariant efficiency) to avoid breaking an implicit etiquette.

Overall, the effects on efficiency gains and losses may be the result of increased or decreased

social pressure to perform depending on the experience and ability of one’s coworker(s). Age is

important to consider, as it is often a determinant of tenure, but given the shape of the

distributions between age and tenure, there is reason to believe that tenure is the more socially

divisible factor.

3.3 Sources of Coworker Effects

The manufacturing firm in this study uses a combination of shift systems, depending on

the plant being observed1. The first is a three-shift system, meaning that every 24 hours there are

three shifts, each eight hours long. In this system, workers do not change their shift time.

Changes in an individual’s composition of coworkers in this system can be attributed to a

combination of factors: vacation time, absenteeism, and open positions caused by turnover all

give workers the chance to fill empty shift positions and earn overtime pay. The second shift

system is a 12-hour system in which workers are assigned to work a combination of four days in

a week. Due to varying shift assignments, a focal worker’s coworkers will change randomly. For

example, if worker A is assigned to work the day shift (7:00am to 7:00pm) on Monday, Tuesday,

Wednesday, and Thursday during a given week, and worker B is assigned Saturday, Sunday,

Monday, and Tuesday during the same shift, workers A and B only work together for two shifts.

However, this could change on a weekly basis, as the schedules for workers A and B change

independently from one another. Because the factors that result in compositional coworker

1 Results by plant can be found in Tables 7, 8, and 9

13

variation are exogenous, it can be assumed that changes are random no matter what shift system

is in use. In addition, work is broken down by department. A worker’s department is determined

by his or her job, so a department consists of only workers with the same role in the production

process. In this paper, a coworker is designated as a worker who works in the same department

during the same shift on a given observation day. There is a total of 19 departments across the

three plants observed. The workers are mostly male, as only 17 out of the 223 observed workers

are female2. Additionally, the typical worker is white/Caucasian - there are no ethnic minorities

in the work environment.

Furthermore, there is no teamwork-centric production in the manufacturing firm. Each

worker is responsible for running his or her machine(s) while working alongside coworkers from

the same work department. During the period of observation, the smallest work team is

comprised of a single worker3, while the largest contained a total of 15 workers. The summary

statistics for team size can be found in Table 7 and the distribution can be found in Figure 3. The

employees’ assigned machines are tied to their department, which only switches on rare

occasions, thus there is no need to control for machine level fixed effects. On the job interactions

can occur between workers but are limited due to the loud noise on the factory floor and the

company policy that enforces the use of earplugs. Coworkers can choose to help each other

should a machine break down, but there is no incentive to do so.

The bulk of worker interaction takes place during the 15-minute breaks and the lunch

period imbedded in the daily work schedule. On breaks, workers tend to associate with the same

groups of coworkers regularly; thus social “cliques” develop, typically compromised of workers

that are familiar with one another. This familiarity can be due to a variety of factors, including

2 Dropping females from the population was shown to have no significant effects on results.

3 Observations in which there are no coworkers present are dropped

14

being in the same shift/department, mutual friends, or pre-existing friendships. During breaks, a

worker can develop an estimation of the work ethic of his coworkers. Due to the layout of the

factory floor and the way machines are oriented it is unlikely that “contemporaneous coworker

effects” (increased productivity caused by a coworker increasing his or her productivity) exist.

Coworkers may observe the output of each other’s machines and can observe how quickly one

another fix problems or load more inputs into the machines; however, it is quite difficult to do so

if a worker is paying undivided attention to his or her own work. Since machines are typically

quite large, workers are often moving about their work area, so coworkers are typically observed

at random intervals. Additionally, since workers typically use multiple machines, it becomes

extremely difficult to judge a coworker’s aggregate output. Productivity can typically only be

judged on the factory floor based on the physical effort put forward; if a coworker is seen

moving quickly and efficiently between machines, it is safe to infer that said coworker can be

judged as a high ability worker, as high efficiency ratings are oftentimes a result of high levels of

effort and skill.

A worker may be influenced by “compositional coworker effects” - changes in

productivity due to the mere presence of a more skilled coworker. A change in a focal worker’s

efficiency could be attributed to a social relationship with a coworker. These effects would likely

not be driven by incentives to outperform one’s peers, due to the lack of incentive pay structures

in the workplace. There are also no formal structures in place to mark highly efficient workers as

high ability performers. Recognition of ability is dependent on a worker’s observations of his or

her coworkers during the production process. Workers can develop a sense of understanding for

a coworker’s average ability through prolonged exposure to said coworker on breaks and during

work periods. Time spent observing the speed and effort of a coworker is valuable in the process

15

of forming an understanding how efficient a given coworker is in the production process.

Moreover, workers may discuss the effort and skill levels of coworkers on their breaks, meaning

that workers likely have superior knowledge of the ability of the coworkers with whom they

socialize. Overall, observing the speed and effort of coworkers and the development of a pool of

common knowledge among in-group coworkers allows workers to identify coworkers with

higher average levels of ability while on the factory floor. For the most part, workers do not

change their departments, but the composition of coworkers on a given work shift does change.

Later, I investigate the method of shift assignment to confirm that it is random and not based on

ability.

While the mechanics of the production process are not teamwork-centric, the business

operations of the firm are. Workers do not generate output unless there is an order from a

customer or the firm wants to increase its inventory by a given amount. Orders are received by

management and passed on to the departments responsible for creating the product that fills the

request. Workers in said departments then work separately such that their aggregate output fulfils

the order. Supervisors are responsible for ensuring that orders are completed with time and

quality standards in mind. Therefore, a worker who underperforms does have a negative effect

on the business since orders cannot be completed as quickly. In short, the production of an

individual is not reliant on his or her peers, but the operations of the firm are heavily reliant on

departments filling given orders.

4. Methodology: Estimating Peer Effects

4.1 Predicting Time Invariant Productivity (Ability) and Proving Random Assignment

16

The first task in the process of measuring coworker effects is predicting “time-invariant

productivity”, referred to as ability. Ability (measured in efficiency, the unit of which is

percentage points) is estimated for each individual worker using the following specification:

(1) Effit = Di + αCjt + γMit + εit

Di is the set of worker fixed effects. Cjt is the set of dummies for every observation day

interacted with department dummies and shift number dummies. Mit is the set of 223 coworker

dummies to control for the presence of a given coworker. Therefore, the dummy variable

“coworker n” is given a value of one if the coworker is in focal worker i’s department and shift,

j, on day, t.

If workers are not randomly assigned to their given departments and shifts, there exists a

possibility that unobservable variables besides changes in work group composition drive the

observed coworker peer effects. The case firm claims that all workers are assigned to their given

jobs randomly and not according to demonstrated ability nor aggregate demand. Shifts are

assigned very far in advance, so changes to worker composition in the face of a large order

cannot occur. Due to a high level of turnover, a new worker can be assigned to any department

and there is no predetermined measure of aptitude for a given job before assignment, so true

ability can only be estimated after training and a sizeable sample of work.

The method for econometrically proving random assignment based on ability (i.e. proof

that high ability workers are not systematically assigned to work with other high ability workers)

follows a similar form to Kato and Shu (2016):

(2) abilityi = βability-i,jt + γability-i,j + αD +εit

abilityi denotes the ability of focal worker i, ability-i,jt is the average ability of worker i’s

coworkers on day t, and ability-i,j is the average ability of all of worker i’s possible coworkers. In

17

other words, any coworker that worked in worker i’s department on the same shift and day

during the length of the eight-month observation period. D denotes weekly fixed effects.

The main coefficient of interest is β, which indicates whether high ability workers are

assigned to work with one another on day t. Because a given worker cannot be assigned as his or

her own coworker, the coworkers are taken from an identical pool of employees that does not

include the focal worker. Therefore, ability-i,j needs to be controlled for because the estimated

value of β would otherwise show an upward bias that becomes more severe with smaller pools

(Guryan et al. 2009 and Kato and Shu 2016).

Table 1 shows evidence in favor of random assignment based on ability. Because the

main variables of interest (ability) are predicted by separate regressions, standard errors shown in

parentheses are bootstrapped with 2000 repetitions and clustered at the individual level. Columns

1 and 2 include the leave-me-out average pool to control for bias in ability-i,jt. Column 1 does not

include week dummies while column two does, showing that controlling for the week of

observation does not affect β. In both columns, β is approaching 0 and is statistically

insignificant. The leave-me-out average is shown to be statistically significant at the 1 percent

level in both cases as expected. Columns 3 and 4 do not include the leave-me-out average and

demonstrate the upward bias on β. Column 3 does not include the dummies controlling for the

week of observation, while Column 4 does.

4.2 Regression Results on Coworker Effects by Tenure

Having confirmed the random assignment of coworkers based on ability, the level

equations and first-difference equations measuring coworker effects can now be established:

(3) Eff = βability-i,jt + πgroupsizejt + α(controls) + εit

(3.1) Eff = µability-i,gjt + γability-g,jt + ρgroupsizeg,jt + τgroupsize-g,jt + α(controls) + εit

18

(4) ΔEff = βΔability-i,jt + Δπgroupsizejt + α(Δcontrols) + εit

(4.1) ΔEff = µΔability-i,gjt + γΔability-g,jt + ρΔgroupsizeg,jt + τΔgroupsize-g,jt + α(Δcontrols) + εit

i denotes a worker, t denotes a day, j denotes a work team (combination of shift and department),

and g denotes a group (temporary or committed). Thus, ability-i,jt is the average ability of worker

i’s coworkers. ability-g,jt is the average ability of worker i’s out-group coworkers and ability-i,gjt is

the average ability of worker i’s in-group coworkers. Controls include hours worked, age, and

department tenure. Gender was not included as a control because the sample size of women is

too small and dropping females was shown to have no significant effect on results. Furthermore,

groupsizejt denotes worker i’s total work team size, groupsizeg,jt denotes worker i’s total number

of in-group coworkers, and groupsize-g,jt denotes worker i’s total number of out-group

coworkers. Additionally, week fixed effects are included in Columns 2 and 4 in Table 2.

Summary statistics for the main variables of interest can also be found in Table 7. Table 2

shows the OLS estimates of equations (4) and (4.1) with bootstrapped standard errors with 2000

repetitions clustered at the individual level. Columns 1 and 2 show no evidence that a focal

worker performs differently when the average ability of coworkers in the work team changes.

This is also the case when weekly fixed effects are included in the regression (column 2). π is not

statistically significant either, meaning that there is no evidence that worker i is influenced by the

size of his or her work team. Columns 3 and 4 show evidence that workers decrease efficiency

according to the ability of their in-group coworkers. µ is statistically significant; a 1 percentage

point increase in the average ability of worker i’s in-group coworkers is associated with an

average decrease of .0262 percentage points in worker i’s efficiency. Therefore, an increase of

one standard deviation in the average ability of worker i’s in-group coworkers is associated with

a 1.39 percentage point decrease in worker i’s efficiency. While this effect is small in magnitude,

it is very consistent. Considering if the coefficients are economically significant depends on how

19

the firm values small fluctuations in efficiency of individuals. While the lower ability workers

may decrease their effort in the presence of high ability workers, it is likely that production of the

high ability worker makes up for the difference.

To check the robustness of the negative peer effect, the same regressions were run with

different tenure thresholds. Tables 4, 5, and 6 display results that are consistent with the results

shown in Table 2. The three supplementary tables all display a negative peer effect with nearly

identical magnitudes, but with the threshold of years of tenure required to be considered a

committed worker being slightly different. Tables 4, 5, and 6 have thresholds of one, three, and

five years, respectively.

4.3 Exploring Interaction Effects and Non-Linearity

Table 3 examines whether being a temporary or committed worker changes the response

to an increase of in-group or out-group ability by adding interaction terms to equations (3) and

(4). Week fixed effects are present in all three OLS specifications. Column 1 shows that neither

committed nor temporary workers differ in their responses to overall increased coworker ability

and that neither group responds significantly different to an increase in work team size. Columns

2 and 3 suggest that neither group has a significantly different response to increased in-group or

out-group ability.

With no significant interaction effect stemming from a single tenure group, it was

deemed possible that the in-group coworker effect could be non-linear. To explore this idea, a

square of average in-group coworker ability and average out-group coworker ability was added

to equation (4.1). The results can be found in Table A4. There is no evidence that the in-group

coworker effect is non-linear, as the square term is statistically insignificant. This result suggests

20

that the relationship between a focal worker’s average daily efficiency and his in-group

coworkers’ ability is more likely to be linear than non-linear.

Additional tests were run to investigate if the observed effects differed based on a focal

worker’s ability relative to the ability distribution of his or her department. There is no evidence

that being above or below the departmental average is associated with an additional response to a

change in average teammate ability. The same can be said for workers in the top and bottom

quartiles of departmental ability distribution. The results can be found in Tables 11 and 12. The

lack of significant results further builds the case that the negative in-group coworker effect is the

same, on average, across the worker population.

5. Results and Discussion

The findings suggest that workers in the focal firm decrease their efficiency when the

average ability of their in-group teammates increases. Moreover, there is no evidence that

changes in the average ability of out-group coworkers has any significant effect. The results

suggest that workers free ride on the higher ability of their in-group coworkers while working

towards a social norm set by plant supervisors. On the other hand, workers are shown to work

harder when the average ability of their similarly tenured coworkers decreases, an effect that I

attribute to a combination of prosocial behavior and increased pressure from managers.

Additionally, the rationale behind the difference in significance between the in and out-group

effects revolves around the relative ease of ability identification among in-group coworkers

compared to out-group coworkers. The lack of any performance incentives in the firm’s wage

structure suggests that workers are not discouraged from or punished for free riding during a

group effort to complete an order.

21

There is straightforward motivation to free-ride when the average ability of one’s

coworkers increases. When an order is given to a department and a focal worker is surrounded

by teammates with above average ability, the focal worker is likely to reduce effort because he or

she knows that the order can be completed without having to extended full effort. Furthermore,

the flat wage structure does not discourage workers from free-riding, as all workers in a given

department are paid the same hourly wage. The effect is reversed when the average ability of

teammates decreases. A focal worker extends a higher than average level of effort when the work

team has a relatively low level of ability. The motivation behind this effect is twofold. The first

is that workers feel a duty to fill the void of ability in the department. For example, a worker may

exhibit prosocial behavior and work harder to complete an order when a high ability teammate is

absent. The second source of motivation comes from supervisors, who will put extra pressure on

a department to complete an order if the average ability of the workers is relatively low that shift

due to an exogenous shock such as sick leave.

The coworker effect is partial to in-group coworkers due to the level of familiarity a focal

worker develops with the other workers in his or her tenure group. Due to the increased exposure

to in-group coworkers, a focal worker has an easier time identifying which of them are low,

average, and high ability. Since workers are less likely to associate with workers in the opposite

group, it becomes difficult to judge ability, as the only exposure to out-group coworkers is on the

factory floor during a work period. Moreover, the effect is impartial to tenure groups, meaning

that neither committed nor temporary workers differ in levels of sensitivity to a change in the

average ability of in-group, out-group, or the whole pool of coworkers.

Additional results show no evidence in favor of the suggestion that the in-group coworker

effect is non-linear, nor that a focal worker’s ability relative to the ability distribution of his or

22

her department is a significant factor in determining the average response to a change in average

coworker ability. The breakdown of the results by plant also offer insight into which plant

populations are driving the effect. The results can be found in tables A1, A2, and A3. Plants 1

and 3 are shown to be the most influential in determining the aggregate effect, which is

unsurprising, given that Plant 2 has a much smaller worker population, so there are fewer

changes in the Plant’s worker pool on a per-shift basis during the period of observation. On a

final note, there was no indication that a change in the size of a work team has any significant

effect on the efficiency of workers within said team.

Overall, the results indicate a small but significant negative coworker effect that is, on

average, consistent across the entire factory population. Higher average in-group coworker

ability motivates workers to free-ride on their coworkers’ efforts towards a social norm set by a

supervisor. The small magnitude and consistency of the effect across the population makes it

difficult for management to prevent free-riding by organizing work teams accordingly. The

implementation of relative performance incentives in the wage structure may erase the incentive

to free ride, but this would be a radical and complicated change for the firm.

6. Conclusion

This paper examines the dynamic between tenure and social identity in the context of

how it effects the efficiency of workers in an economic setting that does not reward performing

better than one’s peers. The setting of an American manufacturing firm provides another

opportunity to study how workers respond to a change in the demographic of their work group.

Social identity theory has suggested that individuals can bond themselves to a group identity

through seemingly insignificant factors, and due to the very uneven distribution of tenure in the

firm, it makes sense that workers bond with those who have similar levels of experience.

23

The evidence suggests that there is an in-group coworker effect; workers decrease their

efficiency to free-ride when the average ability of their similarly tenured coworkers increases.

On the other hand, there is no evidence that workers respond to changes in the average ability of

their out-group coworkers. This is likely due to the relative ease at which a worker can identify

the ability of an in-group coworker compared to an out-group coworker. There is no evidence

that committed or temporary workers respond differently to the coworker effects. Workers are

motivated to free ride by the lack of incentive pay structures, but they are not believed to be

influenced by work team size.

This paper offers insight into a unique empirical setting in which coworker effects are

suggested to exist. Most literature in the realm of peer effects lacks a certain degree of external

validity, so the phenomena shown in this paper should not be considered universal or applicable

to other work settings.

References:

Bandiera, O., Barankay, I., & Rasul, I. (2010). Social incentives in the workplace. The

Review of Economic Studies, 77(2), 417-458.

24

Bernhard, H., Fehr, E., & Fischbacher, U. (2006). Group Affiliation and Altruistic Norm

Enforcement. The American Economic Review, 96(2), 217-221.

Brown, J. (2011). Quitters never win: The (adverse) incentive effects of competing with

superstars. Journal of Political Economy, 119(5), 982-1013.

Chen, Y., & Li, S. X. (2009). Group identity and social preferences. The American Economic

Review, 99(1), 431-457.

Cohn, A., Fehr, E., & Goette, L. (2014). Fair wages and effort provision: Combining

evidence from a choice experiment and a field experiment. Management Science, 61(8),

1777-1794.

Cornelissen, T., Dustmann, C., & Schönberg, U. (2013). Peer effects in the workplace.

Drago, R., & Garvey, G. T. (1998). Incentives for helping on the job: Theory and evidence.

Journal of labor Economics, 16(1), 1-25.

Falk, A., & Ichino, A. (2006). Clean evidence on peer effects. Journal of labor economics,

24(1), 39-57.

Guryan, J., Kroft, K., & Notowidigdo, M. J. (2009). Peer effects in the workplace: Evidence

from random groupings in professional golf tournaments. American Economic Journal:

Applied Economics, 1(4), 34-68.

Haleblian, J., & Finkelstein, S. (1993). Top management team size, CEO dominance, and

firm performance: The moderating roles of environmental turbulence and discretion.

Academy of management journal, 36(4), 844-863.

Kato, T., & Shu, P. (2016). Competition and social identity in the workplace: Evidence from

a Chinese textile firm. Journal of Economic Behavior & Organization, 131, 37-50.

Kaur, S., Kremer, M., & Mullainathan, S. (2010). Self-control and the development of work

arrangements. The American Economic Review, 100(2), 624-628.

Manski, C. F. (1993). Identification of endogenous social effects: The reflection problem.

The review of economic studies, 60(3), 531-542.

Mas, A., & Moretti, E. (2009). Peers at work. The American economic review, 99(1), 112-

145.

Rigotti, L., & Rustichini, A. (2007). Individual behavior and group membership. The

American Economic Review, 97(4), 1340-1352.

Welteke, C. (2015). Peers at Work-a Brief Overview of the Literature on Peer Effects at the

Workplace and the Policy Implications (No. 68). DIW Roundup: Politik im Fokus

Figure 1: Tenure Distribution

25

This figure shows a histogram of each observed worker’s tenure at the firm on the first day of

observation. This shows workers with tenure of four years or less make up a significant

percentage of the worker population.

Figure 2: Age Distribution

26

This figure shows a histogram of each worker’s age on the first day of observation. Comparing

this to the histogram of tenure provides insight into the weak connection between age and tenure

at the firm.

Figure 3: Team Size Distribution

27

This figure is a histogram of the different team sizes during the period of observation. From this,

it can be estimated that workers typically work with between one and six coworkers on any given

day.

28

Table 1: Confirming Random Assignment

This table shows no evidence that workers are assigned based on ability. The dependent variable

is the ability of the focal worker, while the main independent variable is the average ability of

coworkers. The Leave Me Out Average controls for bias.

Table 2: Main Effects

Columns 2 and 4 include weekly fixed effects

Standard errors are bootstrapped with 2000 repetitions and clustered at the individual level

* p<0.05, ** p<0.01, *** p<0.001

Standard errors in parentheses

r2 0.337 0.338 0.0514 0.0541

Observations 11792 11792 11792 11792

(9.030) (9.253) (4.422) (4.450)

Constant 42.86*** 43.15*** 112.1*** 112.5***

(0.0780) (0.0795)

Leave Me Out Average 0.672*** 0.671***

Ability (0.0325) (0.0327) (0.0389) (0.0394)

Teammates' Avg -0.0298 -0.0298 0.0958* 0.0957*

Ability Ability Ability Ability

(1) (2) (3) (4)

Table 1: Confirming Random Assignment

29

This table shows that increased Average In-Group Coworker Ability has a negative effect on a

focal worker’s efficiency. It also shows that the average ability of all teammates has no effect.

The dependent variable is the first difference of efficiency.

Table 3: Interaction Effects



* p<0.05, ** p<0.01, *** p<0.001


Observations 11787 11787 11785 11785

(0.0427) (1.187) (0.0428) (1.199)

Constant 0.00300 4.160*** -0.00247 4.099***

(0.392) (0.396)

FD Out-Group Size 0.344 0.350

(0.593) (0.602)

FD In-Group Size 0.988 0.988

Ability (0.0109) (0.0110)

FD Average Outgroup 0.00440 0.00390

Ability (0.0114) (0.0116)

FD Average In-Group -0.0260* -0.0262*

(0.284) (0.285)

FD Team Size 0.498 0.497

Ability (0.0791) (0.0775)

FD Average Team 0.0165 0.0159

FD Efficiency FD Efficiency FD Efficiency FD Efficiency

(1) (2) (3) (4)

Table 2: Main Effects

30

This table shows that the observed effect shown in Table 2 is not exhibited in a stronger or

weaker magnitude by either committed or temporary workers. The dependent variable is the first

difference of efficiency. No interaction effects are found to be statistically significant.

Table 4: Main Effects, 1 Year Tenure Threshold


* p<0.05, ** p<0.01, *** p<0.001


Observations 11786 11786 11786

(1.209) (1.235) (1.205)

Constant 4.298*** 4.222*** 4.222***

Average Outgroup A~y (0.0403)

Committed * FD -0.0238

Average Ingroup Ab~y (0.0220)

Committed * FD -0.00878

Average Outgroup A~y (0.0402)

Temporary * FD 0.0238

Average Ingroup Ab~y (0.0219)


Ability (0.0108) (0.0392)


Ability (0.0144) (0.0162)

FD Average In-Group -0.0141 -0.00535

Average Team Ability (0.142)


Ability (0.109)

FD Average Team -0.0485

FD Efficiency FD Efficiency FD Efficiency

(1) (2) (3)

Table 3: Interaction Effects

31

This table displays the same elements as Table 2, but in this case the tenure threshold that

determines groups is set to 1 year. The results are nearly identical to those shown in Table 2.




* p<0.05, ** p<0.01, *** p<0.001


Observations 15276 15276 15266 15266

(0.0353) (1.115) (0.0360) (1.108)

Constant 0.00483 3.421** 0.0117 3.461**

(0.333) (0.334)


(0.481) (0.498)


Ability (0.00953) (0.00957)


Ability (0.0100) (0.0104)

FD Average In-Group -0.0261** -0.0261*

(0.287) (0.292)

FD Team Size 0.627* 0.626*

Ability (0.00919) (0.00909)

FD Average Team -0.0226* -0.0227*


(1) (2) (3) (4)


32


determines groups is set to 3 years. The results are nearly identical to those shown in Table 2.




* p<0.05, ** p<0.01, *** p<0.001


Observations 15276 15276 15272 15272

(0.0354) (1.121) (0.0359) (1.101)

Constant 0.00509 3.484** 0.00909 3.409**

(0.317) (0.323)


(0.561) (0.571)


Ability (0.00871) (0.00858)

FD Average Outgroup 0.000143 -0.000345

Ability (0.0102) (0.0103)


(0.255) (0.257)


Ability (0.0109) (0.0110)

FD Average Team -0.0241* -0.0243*


(1) (2) (3) (4)


33


determines groups is set to 5 years. The results are nearly identical to those shown in Table 2.

Table 7: Summary Statistics



* p<0.05, ** p<0.01, *** p<0.001


Observations 15276 15276 15270 15270

(0.0353) (1.119) (0.0357) (1.108)

Constant 0.00508 3.471** 0.00574 3.461**

(0.290) (0.298)


(0.535) (0.548)


Ability (0.00860) (0.00833)


Ability (0.00915) (0.00922)


(0.232) (0.234)


Ability (0.0115) (0.0116)

FD Average Team -0.0260* -0.0263*


(1) (2) (3) (4)


34

N = 11,792

Variable Mean SD Min Max

Age (Years) 43.17 14.044 18.55 66.64

Tenure (Years) 12.69 12.95 0.0027 40.67

Wage ($) 15.17 0.765 14.5 18

Efficiency (% Point) 104.71 27.81 .021 287.33

Average Coworker Ability (% Point) 100.19 32.5 47.07 165.06

Average Coworker Ability In-group (% Point) 93.43 53.02 0 165.06

Average Coworker Ability Out-group (% Point) 49.56 53.93 0 152.72

Team Size 4.32 2.16 2 17

FD Efficiency (% Point) .0024 31.38 -205.69 189.91

FD Average Coworker Ability (% Point) -.002 6.51 -100.68 86.53

FD Average Team Ability Out-group (% Point) -.0055 39.31 -152.72 152.72

FD Average Team Ability In-group (% Point) -.0011 37.48 -165.05 165.05

FD Team Size -.00025 1.67 -12 14

Appendix A: Results by Plant

35

To further investigate the reasoning behind the observed peer effects, the results of

equation (4) were re-run on a by-plant basis. The results can be found in Tables 7, 8, and 9. The

findings suggest that the total coworker and in-group effects are sourced primarily from the

worker populations found in plants 1 and 3. Plant 2’s population does not demonstrate any

significant coworker effect that is of the same sign or magnitude of the overall effect. This is not

too surprising because plant 2’s worker population is the smallest of the three, and therefore the

average day to day change of a focal worker’s coworker population is much smaller than that in

plants 1 and 3.

Table A1: Main Effects, Plant 1




* p<0.05, ** p<0.01, *** p<0.001


Observations 8914 8914 8912 8912

(0.0384) (1.135) (0.0373) (1.087)

Constant -0.000588 3.024** -0.00865 3.111**

(0.382) (0.389)


(0.706) (0.704)

FD In-Group Size -0.134 -0.103

Ability (0.0114) (0.0113)

FD Average Outgroup -0.0124 -0.0125

Ability (0.0113) (0.0111)


(0.274) (0.260)


Ability (0.00990) (0.00974)

FD Average Team -0.0205* -0.0204*


(1) (2) (3) (4)

Table 7: Main Effects, Plant 1

36




* p<0.05, ** p<0.01, *** p<0.001


Observations 2390 2390 2390 2390

(0.0843) (4.715) (0.0852) (4.607)

Constant 0.00134 0.267 0.00113 0.143

(0.956) (1.045)


(2.319) (2.303)

FD In-Group Size -0.434 -0.511

Ability (0.0185) (0.0185)


Ability (0.0333) (0.0339)

FD Average In-Group 0.0171 0.0174

(0.667) (0.696)


Ability (0.0236) (0.0231)

FD Average Team 0.0110 0.0112


(1) (2) (3) (4)


37

Table A4: Checking Non-Linearity



* p<0.05, ** p<0.01, *** p<0.001


Observations 3972 3972 3972 3972

(0.0809) (2.755) (0.0813) (2.795)

Constant -0.00932 5.165 -0.0104 5.248

(1.145) (1.186)


(1.320) (1.334)


Ability (0.0301) (0.0310)

FD Average Outgroup -0.0391 -0.0372

Ability (0.0161) (0.0161)


(0.393) (0.380)


Ability (0.0189) (0.0186)

FD Average Team -0.00244 -0.00220


(1) (2) (3) (4)


38

This table shows that the In-Group coworker effect is likely not non-linear. This OLS estimation

is the same as that shown in Table 2, but with a square of Average Out-Group and In-Group

Ability added. Neither of the square terms are statistically significant.

Appendix B: Laggard Effect and Top/Bottom Quartile Results

Column 2 includes weekly fixed effects


* p<0.05, ** p<0.01, *** p<0.001


Observations 11785 11785

(0.111) (1.172)

Constant 0.0362 4.145***

(0.410) (0.391)


(0.599) (0.611)


Teammate Ability^2 (0.0000493) (0.0000506)

Average Outgroup -0.0000353 -0.0000360

Ability (0.0112) (0.0108)


Teammate Ability^2 (0.0000547) (0.0000570)

Average Ingroup 0.0000113 0.00000929

Ability (0.0113) (0.0115)


FD Efficiency FD Efficiency

(1) (2)

Table 10: Checking Non-Linearity

39

While the interaction effects run in section 4.3 showed no evidence that either group of

workers is more or less sensitive to the coworker effects, it was deemed possible that a focal

worker’s relative ability influences his or her sensitivity to a change in the average ability of

coworkers. To find out if there is evidence behind this idea, OLS estimates of equation (5) were

run.

(5) ΔEff = µΔability-i,gjt + γΔability-g,jt + βΔability-i,gjt *laggard + ρΔgroupsizeg,jt + τΔgroupsize-g,jt + α(Δcontrols) + εit

Equation (5) takes similar form to equation (4.1), but with an interaction term added. The

controls include hours worked, age, and department tenure. The laggard variable is a dummy

variable that takes a value of one if the observed worker has an ability rated lower than the

average of his department, and a value of zero if above. The findings suggest that being below

the departmental average ability is not associated with any additional increase or decrease in the

in-group coworker effect. On average, both above and below average workers respond the same

to an increase in the average ability of in-group coworkers. The results can be found in Table B1.

To investigate whether the coworker effects were truly universal across the ability

distribution, equations (6) and (7) were run.

(6) ΔEff = µΔability-i,gjt + γΔability-g,jt + βΔability-i,gjt *bottomquartile + ρΔgroupsizeg,jt + τΔgroupsize-g,jt + α(Δcontrols) + εit

(7) ΔEff = µΔability-i,gjt + γΔability-g,jt + βΔability-i,gjt *topquartile + ρΔgroupsizeg,jt + τΔgroupsize-g,jt + α(Δcontrols) + εit

These equations are the same as equation (5), but the laggard effect has been replaced by

a dummy that denotes if a coworker is in the bottom or top quartile of the departmental ability

distribution. The results for equations (6) and (7) can be found in columns 1 and 2, respectively,

in Table B2. The findings suggest that there is no significant additional in-group coworker effect

if a focal worker is in the top or bottom quartile of his or her departmental ability distribution.

40

Overall, the findings show that the in-group coworker effect are universal across the

departmental ability distribution, on average. There is no evidence that low ability or below

average ability workers decrease their effort at a higher magnitude than high ability workers

when the average ability of similarly tenured coworkers increases.

Table B1: In-Group Laggard Effect

This table shows an OLS estimation like the Main Effects shown in Table 2, but with an

interaction added. The Laggard dummy interacted with Average In-Group Ability is statistically

insignificant, meaning that there is no evidence that a worker acts differently if he or she is

below the 50th

percentile of the departmental ability distribution.

Column 2 includes weekly fixed effects


* p<0.05, ** p<0.01, *** p<0.001



(0.0421) (1.170)

Constant -0.00208 4.108***

(0.397) (0.398)


(0.605) (0.603)


Ability (0.0110) (0.0111)


Ingroup Teammate A~y (0.0186) (0.0186)

Laggard * Average 0.0163 0.0160

Ability (0.0158) (0.0160)



(1) (2)

Table 11: In-Group Laggard Effect

41

Table B2: In-Group Top/Bottom Quartile Effects

This table shows an OLS estimation like the Main Effects shown in Table 2, but with two

interactions added. The Top and Bottom Quartile dummies are interacted with Average In-Group

Ability, and are statistically insignificant, meaning that there is no evidence that a worker acts

differently if he or she is in the top or bottom quartile of his or her departmental ability

distribution.

Column 2 includes an ingroup * top quartile interaction

Column 1 includes an ingroup * bottom quartile interaction


* p<0.05, ** p<0.01, *** p<0.001



(1.195) (1.184)

Constant 4.105*** 4.105***

Average Ingroup Te~y (0.0237)

Top Quartile * -0.00694

(0.390) (0.402)


(0.592) (0.584)


Ability (0.0111) (0.0109)


Average Ingroup Te~y (0.0210)

Bottom Quartile * 0.0206

Ability (0.0128) (0.0125)

FD Average In-Group -0.0328* -0.0244


(1) (2)

Table 12: In-Group Top/Bottom Quartile Effects

measuring coworker effects: evidence from an american ... · coworkers. productivity can be...

Documents